Research On Co-saliency Detection Via Single-image Saliency Maps

Co-Saliency Object Detection(Co-SOD)aims to discover the commonly salient objects in a group of relevant images,segmenting co-salient objects at pixel-level.As a basic task in the computer vision community,Co-SOD has received more and more attention in recent years.Despite existing data-driven methods gaining promising results on Co-SOD by using deep features,they undervalue the power of intra-saliency cues,leading deep models that cannot achieve satisfying results under both complex and plain scenarios.To this end,this thesis studies the Co-SOD based on the singleimage saliency maps(SISMs),two main contributions are summarized as follows:(1)This thesis introduces SISMs as prior knowledge into deep models,proposing an intra-saliency based Co-SOD model.We first combine SISMs generated by any offthe-shelf SOD model and deep features to produce intra-saliency cues,then exploits consistency in an image group based on such intra-saliency cues to segment co-salient objects.To strengthen the ability of the model on understanding category semantics,we propose self-correlation features to explicitly use category information encoded in the image features,improving the performance of the model for Co-SOD.Experiment results on three benchmark datasets show that the proposed model achieves comparable performance with the best Co-SOD algorithm and robust performance under both complex and plain scenarios,validating the contribution of intra-saliency cues for CoSOD.(2)To better play the role of intra-saliency cues in the Co-SOD,this thesis proposes an intra-saliency-correlation based Co-SOD model.We first propose a correlation fusion module,to explore the correlation between intra-saliency cues from different images,building inter-saliency cues that accurately represent consistency of the image group.Then,we propose a channel-rearranging operation to alleviate the risk of overfitting due to leveraging self-correlation features,further improving the performance of the model.Experiment results show that the proposed model outperforms existing Co-SOD models on three benchmark datasets,ablation studies validate the effectiveness of each designed module and the robustness of the model.